Intelligent Offload Detection for Achieving Approximately Optimal Load Balancing

Abstract

Load balancing is one of the critical issues, which dominates system performance in a storage cluster. To overcome load imbalance, many algorithms have to maintain real-time load conditions that often lead to high-communication overheads. Furthermore, they may require many replicas that significantly decrease storage utilization. In this paper, we present a novel algorithm called intelligent offload detection (IOD) that not only achieves approximately optimal load balancing but also contributes to improve communication overheads and storage utilization. Once non-hotspot data become hotspot data, IOD automatically establishes new partial replicas and replicates them to nodes with minimum load levels in sequence. Next, the arriving requests may be divided into a different number of sub-requests based on the load levels of related nodes. The IOD algorithm works with a partial replica policy so that it is effective to overcome skewed data access originated from hotspot data. Moreover, this policy enhances storage utilization. By classifying the loads into corresponding load levels, IOD is unnecessary for frequent load updates that decrease communication overheads. Simulation results demonstrate that the load balancing in IOD outperforms self-adaptive replication management (SARM) and approaches to optimal load balancing (OLB) under various network conditions. Besides, the amount of required replicas in IOD is relatively smaller than that in SARM and OLB.